OpenCeres: When Open Information Extraction Meets the Semi-Structured Web

Proceedings of NAACL-HLT 2019, pages 3047–3056 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3047 OpenCeres: When Open Information Extraction Meets the Semi-Structured Web Colin Lockard Paul G. Allen School University of Washington lockardc@cs.washington.edu Prashant Shiralkar Amazon shiralp@amazon.com Xin Luna Dong Amazon lunadong@amazon.com Abstract Open Information Extraction (OpenIE), the problem of harvesting triples from natural lan- guage text whose predicate relations are not aligned to any pre-defined ontology, has been a popular subject of research for the last decade. However, this research has largely ignored the vast quantity of facts available in semi- structured webpages. In this paper, we define the problem of OpenIE from semi-structured websites to extract such facts, and present an approach for solving it. We also introduce a labeled evaluation dataset to motivate research in this area. Given a semi-structured website and a set of seed facts for some relations exist- ing on its pages, we employ a semi-supervised label propagation technique to automatically create training data for the relations present on the site. We then use this training data to learn a classifier for relation extraction. Ex- perimental results of this method on our new benchmark dataset obtained a precision of over 70%. A larger scale extraction experiment on 31 websites in the movie vertical resulted in the extraction of over 2 million triples. 1 Introduction Knowledge extraction is the problem of extracting (subject, predicate, object) triples from unstruc- tured or semi-structured data, where the subject and object are entities and the predicate indicates the relationship between them. In conventional in- formation extraction (which we call “ClosedIE”), a closed set of potential predicates and their se- mantics are pre-defined in an ontology. Open In- formation Extraction (OpenIE) is an alternative approach that has no pre-defined ontology and in- stead represents the predicate with a string ex- tracted from the source data. These extractions can capture a much vaster array of semantic relation- ships than ClosedIE and have been used to support many downstream use-cases, including question- answering, ontology discovery, embedding gener- ation, fact checking, and summarization (Mausam, 2016). Previous OpenIE work has concentrated on raw text, with the aim to extract “open” triples from natural language sentences (Niklaus et al., 2018), with another line of work focused on ex- tracting from webtables (Cafarella et al., 2008). Semi-structured websites (e.g. IMDb) contain many pages displaying information in stand-alone fields in relatively consistent locations on each page, with entities and the relationship between them indicated via formatting features such as section headers and lists of key-value pairs. Fig- ure 1 shows an example page and the triples it conveys. Semi-structured websites have recently been shown to be a rich target for IE; the Knowl- edge Vault large-scale web extraction experiment, which extracted from semi-structured websites, natural language text, webtables, and Semantic Web annotations, found that semi-structured web- sites contributed 75% of total extracted facts and 94% of high-confidence extractions (Dong et al., 2014); the Ceres system showed that one can au- tomatically …

Content truncated to 500 words